A dough mixing machine control method based on a one-dimensional convolutional neural network and a dough mixing machine

Abstract

The application discloses a dough mixing machine control method and dough mixing machine based on one-dimensional convolutional neural network and belongs to the field of intelligent light industry machinery control. The method first collects static parameters such as flour weight, water weight, initial temperature and humidity, motor rated power and torque constant, then collects dynamic time series data such as motor running current, real-time speed and real-time temperature and humidity in the flour cylinder, fuses multi-source data to construct a 7-channel time series feature matrix. One-dimensional convolution kernel covering all channels is used for feature extraction, rheology correction is performed through temperature and humidity time series features, and one-dimensional convolutional neural network model deployed on the STM32 main controller is used to infer dough forming stage probability in real time. The application can adapt to the reference offset caused by different feeding amounts, flour-water ratios and motor hardware differences, improve the accuracy of dough mixing endpoint judgment under different environmental conditions, and reduce the probability of false shutdown caused by single-point data fluctuation.

Description

Technical Field

[0001] This invention relates to the field of intelligent light industrial machinery control technology, and in particular to a dough mixer control method and a dough mixer based on a one-dimensional convolutional neural network. Background Technology

[0002] In the field of industrial and household dough mixers, judging the dough's kneading state is a crucial step affecting dough quality. Current technologies primarily use the following methods to determine the kneading endpoint: Firstly, there is the timed program control method, which controls the dough kneading process by preset a fixed duration. This method cannot cope with the differences in gluten formation speed caused by different flour water absorption rates and changes in environmental temperature and humidity, making it difficult to guarantee the shaping quality of the dough under different working conditions.

[0003] Secondly, the single current or torque curve method monitors changes in motor current and stops kneading when the current reaches its peak or a stable state. This method does not consider the initial total mass of flour and water, or the influence of the flour-to-water ratio on the current reference. Different amounts of flour will result in different absolute current values ​​under the same kneading conditions, which can easily lead to premature stopping before reaching the optimal shaping state or over-kneading causing gluten breakage.

[0004] Third, existing control methods do not fully consider the impact of ambient temperature differences. The formation of the gluten network is a thermosensitive process, and the rheological characteristics of the dough will shift under different room temperature conditions. Traditional equipment has difficulty maintaining consistent judgment accuracy in different seasons.

[0005] Therefore, further improvements are needed. Summary of the Invention

[0006] The purpose of this invention is to address the shortcomings of existing technologies by providing a dough mixer control method and a dough mixer based on a one-dimensional convolutional neural network. Through multi-source parameter fusion and edge artificial intelligence inference, the accuracy of dough mixer endpoint determination is improved, and the invention is adaptable to different feed amounts, flour-to-water ratios, and environmental temperature and humidity conditions.

[0007] One embodiment of the present invention addresses its technical problem by employing the following technical solution: a dough mixer control method based on a one-dimensional convolutional neural network, comprising the following steps: S1: Static parameter acquisition steps: Before the dough mixing begins, collect the weight of the flour and water added, and collect the initial temperature and humidity inside the dough container; obtain the parameters of the current dough mixer motor, including the rated power and torque constant; S2: Dynamic data acquisition steps: Start the motor to knead the dough. During the mixing process, multi-dimensional dynamic time-series characteristic data is collected in real time at a preset sampling frequency. The multi-dimensional dynamic time-series characteristic data includes motor operating current, motor real-time speed and real-time temperature and humidity inside the dough cylinder. S3: Feature matrix construction steps: The flour weight, water weight, initial temperature and humidity, and motor parameters are fused with the multi-dimensional dynamic time series feature data to construct a multi-channel time series feature matrix; S4: Real-time state reasoning step: Input the multi-channel temporal feature matrix into a one-dimensional convolutional neural network model that has been pre-trained and quantized and deployed in the main controller. The model calculates and outputs the probability values ​​of the dough being in different forming stages in real time. S5: Closed-loop shutdown control steps: When the probability value of the optimal forming stage output by the one-dimensional convolutional neural network model is greater than or equal to the preset confidence threshold, and the duration reaches the preset time threshold, the main controller controls the motor to stop running.

[0008] Further, in step S3, the multi-channel timing feature matrix is ​​a two-dimensional matrix with 7 channels × L timing length, where L is the number of sampling points corresponding to the sliding window; the 7 channels are motor operating current, motor real-time speed, real-time dough temperature, initial flour weight, initial water weight, motor rated power, and motor torque constant, where the static constant and solidified parameters are aligned with the dynamic timing data on the time axis by broadcasting and copying L times.

[0009] Furthermore, in step S4, the one-dimensional convolutional neural network uses a one-dimensional convolutional kernel that vertically covers all 7 channels. During the sliding process, the convolutional kernel enables the network to nonlinearly couple static boundary conditions with dynamic waveforms at the bottom layer, adaptively eliminating the reference current offset caused by different operating conditions.

[0010] Furthermore, in step S4, the one-dimensional convolutional neural network model undergoes rheological correction using time-series temperature and humidity features, specifically including: The initial humidity collected is used to correct the equilibrium moisture content of the flour's own environment, thereby reducing the kinetic response error caused by changes in the flour's water absorption rate. By extracting the first-order local rate of change of the real-time dough temperature sequence through lateral convolution, the slope inflection characteristics of the dough temperature rise curve are identified, forming a synergistic criterion with the motor current characteristics.

[0011] Furthermore, in step S5, the main controller internally runs dynamic smoothing filtering logic based on a finite state machine, which includes four operating states: Mixing and stirring state: During normal stirring, if the probability of the best forming stage of single frame inference output is greater than 90%, the state machine judges it as occasional impact noise and does not perform state transition. Pre-forming state confirmation: When the probability of the periodic inference output of the one-dimensional convolutional neural network in the optimal forming stage is greater than or equal to 90%, the confirmation mechanism is triggered, and the main controller opens a sliding time window queue with a length of 4 in the internal memory. Safe stop state: If the average probability of all inference cycles in the queue is greater than or equal to 92% within the next 4 consecutive seconds, the state machine will jump to the safe stop state and the main controller will control the motor to stop running. Over-kneading interruption state: If the probability of over-kneading stage output by the one-dimensional convolutional neural network model is greater than or equal to 85% at any stage, the system skips the sliding window counting and directly triggers the highest priority interrupt to perform a shutdown operation.

[0012] Furthermore, the main controller uses an STM32 series chip; the weighing acquisition uses an HX711 weighing module; and the temperature and humidity acquisition uses an SHT30 temperature and humidity sensor.

[0013] Furthermore, in step S4, the training and deployment process of the one-dimensional convolutional neural network model includes: We obtained different flour weights, water weights, motor parameters, and corresponding dynamic operating current, speed, temperature, and humidity time-series samples through offline testing, and labeled them according to the state of the dough to build a training dataset. The one-dimensional convolutional neural network model is iteratively trained on a computer using the training dataset. The trained model is reduced by 8-bit integer quantization to generate a lightweight model; The lightweight model is converted into C language code using a model conversion tool and then burned and deployed to the Flash memory of the main controller.

[0014] Furthermore, in step S2, the preset sampling frequency is 20Hz, and the dynamic operating current and the real-time speed of the motor are collected in alignment on the same hardware microsecond-level timestamp.

[0015] On the other hand, the present invention provides a dough mixer, including a main controller, a weighing module, a temperature and humidity sensor, a motor drive module, and a computer program stored in the main controller and executable on the main controller; when the computer program is executed by the main controller, the main controller performs the dough mixer control method based on a one-dimensional convolutional neural network as described in any one of claims 1 to 8.

[0016] Furthermore, the weighing module includes an HX711 weighing module and a weighing sensor, the weighing sensor being located at the bottom of the dough cylinder; the temperature and humidity sensor is an SHT30 temperature and humidity sensor with a waterproof and dustproof PE sintered shell, fixed at the bottom edge of the dough cylinder, and in close contact with the heat-conducting metal sheet on the inner wall of the dough cylinder through thermally conductive silicone grease.

[0017] The beneficial effects of this invention are as follows: By fusing static material characteristics, motor hardware parameters, and dynamic time-series feedback data from multiple sources, this invention constructs a 7-channel time-series feature matrix. Feature extraction is performed using a one-dimensional convolutional kernel that vertically covers all channels. This allows the network to directly achieve nonlinear coupling between static boundary conditions and dynamic waveforms at the underlying layer, adaptively eliminating reference current offsets caused by differences in feed amounts, flour-to-water ratios, and motor hardware. Rheological correction is achieved by introducing temperature and humidity time-series features. Initial humidity is used to correct for differences in flour water absorption. By extracting the slope transition features of the temperature sequence and forming a collaborative criterion with current features, the influence of environmental factors on dough rheological characteristics can be reduced, improving the accuracy of dough endpoint judgment under different seasons and temperature / humidity conditions. A quantized, lightweight one-dimensional convolutional neural network model is deployed on an embedded STM32 main controller, enabling real-time inference at the edge, with fast response speed, eliminating reliance on cloud computing, and reducing system latency and network dependence. The main controller internally runs dynamic smoothing filtering logic based on a finite state machine. By using a sliding time window to perform mean filtering on the inference probability, it can reduce the probability of false shutdowns caused by single-point probability fluctuations due to noise in the industrial field. At the same time, an over-kneading over-fuse protection mechanism is set up, which can respond quickly when the risk of over-kneading of dough is detected, reducing the occurrence of gluten breakage.

[0018] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0019] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which: Figure 1 This is a block diagram of the hardware structure of the system of the present invention; Figure 2 This is a flowchart of the STM32 closed-loop anti-accidental touch control program of the present invention. Detailed Implementation

[0020] This section will describe in detail specific embodiments of the present invention. Preferred embodiments of the present invention are shown in the accompanying drawings. The purpose of the drawings is to supplement the textual description with graphics, so that people can intuitively and vividly understand each technical feature and overall technical solution of the present invention, but they should not be construed as limiting the scope of protection of the present invention.

[0021] In the description of this invention, "multiple" means two or more; "greater than," "less than," and "exceeding" are understood to exclude the stated number; "above," "below," and "within" are understood to include the stated number. The use of "first" and "second" in the description is merely for distinguishing technical features and should not be construed as indicating or implying relative importance, or implicitly indicating the number of indicated technical features, or implicitly indicating the order of the indicated technical features.

[0022] In the description of this invention, it should be understood that the orientation descriptions, such as up, down, front, back, left, right, etc., are based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limiting this invention.

[0023] In this invention, unless otherwise explicitly defined, the terms "setting," "installing," and "connecting" should be interpreted broadly. For example, they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to a fixed connection, a detachable connection, or an integrally formed connection; they can refer to a mechanical connection; they can refer to the internal connection of two components or the interaction between two components. Those skilled in the art can reasonably determine the specific meaning of the above terms in this invention in conjunction with the specific content of the technical solution.

[0024] Example Reference Figure 1 and Figure 2 , Figure 1 The block diagram centers on the STM32 main controller; the left input terminal connects to the HX711 weighing module and the weighing sensor at the bottom of the face cylinder, as well as the SHT30 temperature and humidity sensor; the upper input terminal connects to the motor current acquisition circuit and the speed sampling sensor; the right control output terminal connects to the motor driver and the buzzer prompt unit. Figure 2 The flowchart of the STM32 closed-loop anti-accidental touch control program shows the process from initial static weighing and temperature / humidity readings after startup, to 20Hz dynamic timing sampling and feature matrix construction, to quantized 1D-CNN inference layer probability calculation, to a sliding window check counter to determine if the threshold is met for 4 consecutive seconds, and finally to either "keep stirring" or "interrupt trigger shutdown". The present invention proposes a dough mixer control method and dough mixer based on a one-dimensional convolutional neural network, which achieves precise control of the dough mixing endpoint through multi-source parameter fusion and edge AI linkage.

[0025] 1. System hardware topology and sampling settings: This system uses an STM32 controller as its core. A load cell is embedded in the bottom of the dough container, working in conjunction with an HX711 weighing module to collect the initial static mass of flour and water before kneading begins. An SHT30 temperature and humidity sensor, with a waterproof and dustproof PE sintered shell, is fixedly positioned at the bottom edge of the stationary dough container. It is in close contact with the thermally conductive metal sheet on the inner wall of the dough container via thermal grease, and is used to collect initial and dynamic temperature and humidity data.

[0026] The motor's drive circuit includes an on-chip phase current amplification module and a speed pulse shaping module to provide high-fidelity dynamic characteristics. For current sampling, a milliohm-level precision sampling resistor is connected in series at the low end of the power switch. The current flowing through it generates a weak voltage drop, which is amplified by a high-speed differential operational amplifier and filtered by an RC low-pass filter to remove high-frequency switching noise before being connected to the high-precision analog-to-digital converter pin of the STM32. For speed sampling, a Hall effect sensor embedded in the motor is used. As the motor rotor rotates, the Hall effect sensor outputs alternating high and low level pulses. After pulse signal shaping by pull-up resistors and Schmitt triggers on the driver board, the pulses are introduced to the timer pin of the STM32 with input capture functionality.

[0027] The STM32 main controller uses a timer configured in input capture mode to latch the speed pulse period in real time to calculate the speed. Simultaneously, it uses the midpoint of the pulse width modulation complementary output to trigger a high-precision analog-to-digital converter to acquire current data. This setting aligns the dynamic operating current and the motor's real-time speed on the same hardware microsecond-level timestamp; in this embodiment, the sampling frequency is set to 20Hz.

[0028] 2. Construction of multi-channel time series feature matrix: The input to the one-dimensional convolutional neural network is a two-dimensional matrix with 7 channels × L time sequence length. In this embodiment, the sliding window is 5 seconds, so L=100. The alignment of the 7 channels at each sampling point t on the time axis is shown in the table below:

[0029] Convolutional kernel fusion mechanism: The one-dimensional convolutional kernel of the one-dimensional convolutional neural network covers 7 channels vertically. During the sliding process, the network can directly nonlinearly couple the static boundary conditions with the dynamic waveform at the bottom layer, thereby adaptively eliminating the reference current offset caused by different operating conditions.

[0030] 3. Rheological correction mechanism of temperature and humidity time series, including effective surface water ratio compensation and temperature rise slope determination method.

[0031] Effective flour-to-water ratio compensation: The initial humidity collected by SHT30 is used to correct the equilibrium moisture content of the flour's own environment, eliminating the kinetic response error caused by changes in the flour's water absorption rate.

[0032] Temperature rise slope determination method: During the mixing process, the dough temperature rise curve T(t) is strongly correlated with the gluten expansion stage. In the unformed stage, mechanical resistance is high, and the temperature rise exhibits a steep first slope; when the dough reaches the optimal forming state (kneaded well), the dough surface becomes smooth, internal energy is converted into network structure tension, and friction with the mixing drum is in dynamic equilibrium, the temperature rise curve slows down, exhibiting a second slope (the first slope is greater than the second slope). 1D-CNN automatically extracts the first-order local rate of change (slope inflection) of the temperature sequence through lateral convolution, forming a co-criterion with the current peak feature.

[0033] Training and deployment of a one-dimensional convolutional neural network model: First, a training dataset is built through offline testing, covering different flour weights, water weights, motor parameters, and different environmental temperature and humidity conditions. The corresponding dynamic operating current, speed, temperature and humidity time series samples are collected, and professionals label the state according to the actual forming state of the dough, including three stages: unformed, optimally formed, and over-kneaded.

[0034] On the computer, the aforementioned training dataset is used to iteratively train a one-dimensional convolutional neural network, optimizing the network parameters to enable the model to accurately identify the dough forming state under different working conditions. After training, the model is reduced by 8-bit integer quantization, significantly reducing the number of model parameters and computational cost while maintaining inference accuracy, generating a lightweight model suitable for embedded platforms.

[0035] Finally, the lightweight model was converted into C language code using a model conversion tool and burned into the Flash memory of the STM32 main controller to achieve real-time inference at the edge.

[0036] 4. STM32 internal anti-accidental touch logic and state machine design: To reduce the possibility of accidental shutdown caused by single-point probability fluctuations due to noise in industrial environments, the main controller internally runs dynamic smoothing filtering logic based on a finite state machine, specifically including the following four operating states: Mixing and stirring state: The system enters this state after startup and performs normal stirring operations. In this state, even if the probability of the optimal forming stage of a single frame inference output occasionally exceeds 90%, the state machine determines it as occasional impact noise, does not perform a state transition, and continues stirring.

[0037] Pre-forming state confirmation: When the one-dimensional convolutional neural network performs periodic inference at a frequency of 1Hz and outputs an optimal forming stage probability greater than or equal to 90%, the pre-forming confirmation mechanism is triggered. The main controller allocates a sliding time window queue of length 4 in its internal memory to store the probability values ​​of the subsequent four consecutive inference cycles.

[0038] Safe Stop State: If, within the next 4 consecutive seconds, the average probability of all inference cycles in the sliding time window queue is greater than or equal to 92%, the state machine transitions to the safe stop state, the main controller immediately cuts off the pulse width modulation signal of the motor driver, and the motor performs a stop operation. If this condition is not met, the state machine returns to the mixing and stirring state and continues mixing.

[0039] Over-kneading trigger state: During the inference process, the model outputs the probability of the over-kneading stage. If the probability of the over-kneading stage is greater than or equal to 85% at any stage of operation, it indicates that there is a risk of gluten breakage in the dough. The system skips the sliding window counting process and directly triggers the highest priority interrupt to perform a shutdown operation.

[0040] Of course, the present invention is not limited to the above-described embodiments. Those skilled in the art can make equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications and substitutions are included within the scope defined by the claims of this application.

Claims

1. A dough mixer control method based on a one-dimensional convolutional neural network, characterized in that, Includes the following steps: S1: Static parameter acquisition steps: Before the dough mixing begins, collect the weight of the flour and water added, and collect the initial temperature and humidity inside the dough container; obtain the parameters of the current dough mixer motor, including the rated power and torque constant; S2: Dynamic data acquisition steps: Start the motor to knead the dough. During the mixing process, multi-dimensional dynamic time-series characteristic data is collected in real time at a preset sampling frequency. The multi-dimensional dynamic time-series characteristic data includes motor operating current, motor real-time speed and real-time temperature and humidity inside the dough cylinder. S3: Feature matrix construction steps: The flour weight, water weight, initial temperature and humidity, and motor parameters are fused with the multi-dimensional dynamic time series feature data to construct a multi-channel time series feature matrix; S4: Real-time state reasoning step: Input the multi-channel temporal feature matrix into a one-dimensional convolutional neural network model that has been pre-trained and quantized and deployed in the main controller. The model calculates and outputs the probability values ​​of the dough being in different forming stages in real time. S5: Closed-loop shutdown control steps: When the probability value of the optimal forming stage output by the one-dimensional convolutional neural network model is greater than or equal to the preset confidence threshold, and the duration reaches the preset time threshold, the main controller controls the motor to stop running.

2. The dough mixer control method based on a one-dimensional convolutional neural network according to claim 1, characterized in that: In step S3, the multi-channel timing feature matrix is ​​a two-dimensional matrix with 7 channels × L timing length, where L is the number of sampling points corresponding to the sliding window; the 7 channels are motor operating current, motor real-time speed, real-time dough temperature, initial flour weight, initial water weight, motor rated power, and motor torque constant, respectively. The static constant and solidified parameters are aligned with the dynamic timing data on the time axis by broadcasting and copying L times.

3. The dough mixer control method based on a one-dimensional convolutional neural network according to claim 2, characterized in that: In step S4, the one-dimensional convolutional neural network uses a one-dimensional convolutional kernel that covers all 7 channels vertically. During the sliding process, the convolutional kernel enables the network to nonlinearly couple static boundary conditions with dynamic waveforms at the bottom layer, adaptively eliminating the reference current offset caused by different operating conditions.

4. The dough mixer control method based on a one-dimensional convolutional neural network according to claim 1, characterized in that: In step S4, the one-dimensional convolutional neural network model undergoes rheological correction using time-series temperature and humidity features, specifically including: The initial humidity collected is used to correct the equilibrium moisture content of the flour's own environment, thereby reducing the kinetic response error caused by changes in the flour's water absorption rate. By extracting the first-order local rate of change of the real-time dough temperature sequence through lateral convolution, the slope inflection characteristics of the dough temperature rise curve are identified, forming a synergistic criterion with the motor current characteristics.

5. The dough mixer control method based on a one-dimensional convolutional neural network according to claim 1, characterized in that: In step S5, the main controller internally runs dynamic smoothing filtering logic based on a finite state machine, which includes four operating states: Mixing and stirring state: During normal stirring, if the probability of the best forming stage of a single frame inference output is greater than 90%, the state machine judges it as occasional impact noise and does not perform state transition. Pre-forming state confirmation: When the probability of the periodic inference output of the one-dimensional convolutional neural network in the optimal forming stage is greater than or equal to 90%, the confirmation mechanism is triggered, and the main controller opens a sliding time window queue with a length of 4 in the internal memory. Safe stop state: If the average probability of all inference cycles in the queue is greater than or equal to 92% within the next 4 consecutive seconds, the state machine will jump to the safe stop state and the main controller will control the motor to stop running. Over-kneading interruption state: If the probability of over-kneading stage output by the one-dimensional convolutional neural network model is greater than or equal to 85% at any stage, the system skips the sliding window counting and directly triggers the highest priority interrupt to perform a shutdown operation.

6. The dough mixer control method based on a one-dimensional convolutional neural network according to claim 1, characterized in that: The main controller uses an STM32 series chip; the weighing acquisition uses an HX711 weighing module; and the temperature and humidity acquisition uses an SHT30 temperature and humidity sensor.

7. The dough mixer control method based on a one-dimensional convolutional neural network according to claim 1, characterized in that: In step S4, the training and deployment process of the one-dimensional convolutional neural network model includes: We obtained different flour weights, water weights, motor parameters, and corresponding dynamic operating current, speed, temperature, and humidity time-series samples through offline testing, and labeled them according to the state of the dough to build a training dataset. The one-dimensional convolutional neural network model is iteratively trained on a computer using the training dataset. The trained model is reduced by 8-bit integer quantization to generate a lightweight model; The lightweight model is converted into C language code using a model conversion tool and then burned and deployed to the Flash memory of the main controller.

8. The dough mixer control method based on a one-dimensional convolutional neural network according to claim 1, characterized in that: In step S2, the preset sampling frequency is 20Hz, and the dynamic operating current and the real-time speed of the motor are collected in alignment on the same hardware microsecond-level timestamp.

9. A dough mixer, characterized in that, It includes a main controller, a weighing module, a temperature and humidity sensor, a motor drive module, and a computer program stored in the main controller and capable of running on the main controller; when the computer program is executed by the main controller, it causes the main controller to perform the dough mixer control method based on a one-dimensional convolutional neural network as described in any one of claims 1 to 8.

10. The dough mixer according to claim 9, characterized in that: The weighing module includes an HX711 weighing module and a weighing sensor, with the weighing sensor located at the bottom of the dough cylinder; the temperature and humidity sensor is an SHT30 temperature and humidity sensor with a waterproof and dustproof PE sintered shell, fixed at the bottom edge of the dough cylinder, and in close contact with the heat-conducting metal sheet on the inner wall of the dough cylinder through thermal grease.

Images